Trillboards OpenRTB 2.6 Extension Schema Reference

version 2026-03-v1

Schema reference for Trillboards' proprietary OpenRTB 2.6 bid request extensions. This is NOT a bid endpoint specification — it documents the SIGNALS injected into every bid request sent to DSPs so engineers can build bidding algorithms that leverage Trillboards' unique audience intelligence.

How to Use This Document

When Trillboards sends a bid request to your DSP, it includes standard OpenRTB 2.6 fields plus the proprietary extensions documented here. Your bidding algorithm can use these signals to bid higher on high-value inventory.

Extension Namespaces

Signal Sources

• FEIN Edge AI: On-device ML Kit (face detection, gaze tracking) + YAMNet TFLite (audio classification) on the tablet, which sends Trillboards a record for each detected person and sends selected camera frames to a cloud model.
• VAS Engine: Verified Attention Seconds computed from historical FEIN measurements, stored in PostgreSQL attention_ledger.
• Intent Graph: Speech-derived purchase intent from YAMNet audio scene classification + keyword detection on the tablet. The tablet also sends Trillboards the speech transcript, which a cloud model reads for purchase intent and brands mentioned.
• Weather Service: WeatherAPI.com data cached in Redis with 60-minute TTL.
• Census ACS: US Census American Community Survey neighborhood demographics by geohash.

Data behind the signals

• Bid requests carry the screen's IP address, its location (its live GPS position where the screen reports one) and its live audience measurements.
• The tablet keeps a record for each person its camera detects: estimated age range, gender, dominant emotion, dwell time and the time spent looking at the screen. On tablets running the full sensing profile it also computes a face-identity template, which Trillboards uses to recognise the same person across screens and days.
• Selected camera frames go to a cloud model (Google Vertex AI) that estimates ethnicity, age band, gender, social composition, purchase-intent stage, activity, attire, carried items, lifestyle segments and mood. Trillboards keeps those frames for 400 days; the frames captured between 5 May and 21 June 2026 are kept as a model-evaluation set with no scheduled deletion.
• Audio: the tablet classifies ambient sound and transcribes speech on the device, and Trillboards stores the transcript. No audio recording is stored or transmitted.
• Radio: the tablet sends Trillboards the Bluetooth, Wi-Fi and local-network identifiers nearby phones broadcast. Trillboards stores them and links device keys to the person seen with them and to mobile advertising IDs.
• Trillboards sells and shares this audience measurement and location data with its data partners, including 375ai and Intuizi. Each category of data partner and what it receives: https://trillboards.com/legal/data-partners

Schemas

30 schemas. Every one is derived from docs/openapi/openrtb-extensions.yaml at request time.

AudienceAttentionPrediction · AudienceSocialAttention · BidRequestExt · CompletedImpressions · ContractExt · DOOHBidRequest · DOOHObject · DemandPartners · Device · EarnerScreens · EventExt · FootTrafficExt · IABSegmentReference · Impression · IntentExt · OpenrtbRevenueScreenHourly · PartnerReconciliationReports · PartnerVariants · PartnerZoneCpmDaily · Regs · Segment · SellerDefinedSegmentReference · Source · TrbNamespace · TrillboardsAudienceExt · User · UserData · VASExt · VenueExt · WeatherExt

AudienceAttentionPrediction

Backend-derived predictive attention signal + buyer-grade attentive-seconds. 1:1 with audience_metrics. Never null by contract; placeholder on row 1.

AudienceSocialAttention

Buyer-grade social amplification signal (5 typed columns). Cloud Vertex Gemini producer. 1 row per audience observation when cascade is observable.

BidRequestExt

Top-level bid request extensions.

⚠️ Both Trillboards namespaces are STRIPPED on curated (registered-audience) deals

When a bid request is served under a curation plan with registeredAudienceOnly — the shape behind our curated PMP / registered audience deals — the server deletes ext.trillboards and ext.trb outright before the request leaves us (services/programmatic/bidstreamSignalCuration.js). If ext is left with no other keys, ext itself is removed.

This is deliberate: on a curated deal the audience is carried by the registered segment in user.data, and the raw sensing signals are withheld.

The consequence for a bidder is the trap worth stating plainly: every signal described on this page disappears on exactly the deals we most want you buying. A bidder whose valuation depends on ext.trb.vas, ext.trillboards.attention_score, or any other field here will price curated inventory as though the screen had no audience at all.

Build for this explicitly:

• Treat both namespaces as absent-by-default and gate every read on their presence. Do not assume a partially-populated ext.
• Do not infer "no audience" or a zero-valued audience from their absence.
• On a curated deal, read the audience from user.data (the registered segment) instead. That is where the targeting signal lives for these deals.
• The stripping is all-or-nothing per request: you will never see ext.trb survive while ext.trillboards is removed, or vice versa.

CompletedImpressions

Completed-play observations from VAST complete and proof-of-play paths. Deduped by issued request/impression key for live revenue and daily earnings.

ContractExt

Signal quality assurance contract. Tells DSPs how complete and reliable the signals in this bid request are. If valid=false, some required fields are missing (but the request is still valid OpenRTB).

{
  "version": "2026-03-v1",
  "valid": true,
  "required_field_coverage_pct": 100,
  "error_count": 0,
  "warning_count": 0,
  "errors": [],
  "warnings": []
}

DOOHBidRequest

Full OpenRTB 2.6 DOOH bid request with all Trillboards extensions

{
  "id": "691fced1-040c-4354-b861-ef2496f44f29",
  "imp": [
    {
      "id": "imp_507f1f77bcf86cd799439011",
      "tagid": "507f1f77bcf86cd799439011",
      "bidfloor": 2,
      "bidfloorcur": "USD",
      "secure": 1,
      "dt": 1788279423607,
      "qty": {
        "multiplier": 2,
        "sourcetype": 2
      },
      "banner": {
        "w": 1920,
        "h": 1080,
        "format": [
          {
            "w": 1920,
            "h": 1080
          }
        ],
        "pos": 1
      },
      "video": {
        "mimes": [
          "video/mp4",
          "video/webm"
        ],
        "minduration": 5,
        "maxduration": 30,
        "protocols": [
          2,
          3,
          5,
          6
        ],
        "w": 1920,
        "h": 1080,
        "startdelay": 0,
        "linearity": 1,
        "skip": 0,
        "placement": 4
      }
    }
  ],
  "dooh": {
    "id": "507f1f77bcf86cd799439011",
    "name": "Test Screen",
    "venuetype": [
      2,
      202
    ],
    "venuetypetax": 5,
    "publisher": {
      "id": "trillboards",
      "name": "Trillboards",
      "domain": "trillboards.com",
      "cat": [
        "IAB24"
      ]
    },
    "domain": "trillboards.com",
    "keywords": "Retail,Shopping Mall,New York,NY",
    "content": {
      "genre": "Entertainment",
      "cattax": 1,
      "cat": [
        "IAB1"
      ],
      "prodq": 1,
      "context": 1,
      "contentrating": "TV-PG",
      "livestream": 1,
      "language": "EN",
      "channel": {
        "name": "Entertainment"
      }
    }
  },
  "device": {
    "devicetype": 8,
    "connectiontype": 2,
    "w": 1920,
    "h": 1080,
    "geo": {
      "lat": 40.7128,
      "lon": -74.006,
      "type": 1,
      "country": "US",
      "region": "NY",
      "city": "New York",
      "zip": "10001"
    }
  },
  "source": {
    "fd": 0,
    "tid": "691fced1-040c-4354-b861-ef2496f44f29",
    "schain": {
      "complete": 1,
      "nodes": [
        {
          "asi": "trillboards.com",
          "sid": "trillboards",
          "rid": "691fced1-040c-4354-b861-ef2496f44f29",
          "name": "Trillboards",
          "domain": "trillboards.com",
          "hp": 1
        }
      ],
      "ver": "1.0"
    }
  },
  "at": 1,
  "tmax": 500,
  "cur": [
    "USD"
  ],
  "test": 0,
  "regs": {
    "coppa": 0,
    "gdpr": 0
  },
  "ext": {
    "protocol": "OpenRTB 2.6",
    "trillboards_contract": {
      "version": "2026-03-v1",
      "valid": true,
      "required_field_coverage_pct": 100,
      "error_count": 0,
      "warning_count": 0,
      "errors": [],
      "warnings": []
    },
    "trillboards": {
      "attention_score": 0.87,
      "audience_count": 8,
      "crowd_density": 12,
      "ad_receptivity": 0.75,
      "emotional_engagement": 0.82,
      "mood": "happy",
      "estimated_occupancy": 25,
      "avg_dwell_ms": 4500,
      "group_composition": {
        "solo": 3
      },
      "screen_engagement": "active",
      "audience_reaction": "positive",
      "signals_age_ms": 12480
    },
    "trb": {
      "weather": {
        "temp_f": 85,
        "feels_like_f": 88,
        "humidity": 75,
        "condition": "Partly cloudy",
        "uv_index": 8,
        "wind_mph": 12,
        "is_raining": false,
        "temp_segment": "hot"
      },
      "foot_traffic": {
        "current_audience": 25,
        "estimated_occupancy": 50,
        "avg_dwell_seconds": 6,
        "traffic_tier": "very_high",
        "estimated_hourly_impressions": 15000
      }
    }
  }
}

DOOHObject

OpenRTB 2.6 DOOH object (top-level, replaces site/app)

DemandPartners

Canonical INT-PK identity table for demand-side partners (DSPs/SSPs that Trillboards routes ad requests to). Wave 2 of pristine-state refactor. Replaces partner_key TEXT joins across the billing path.

Device

EarnerScreens

User-owned ad screens. Each row is a logical screen that can display ads. Links user to device. PG primary with Mongo legacy mirror.

EventExt

Large-event proximity for the screen. The whole event sub-namespace is present ONLY when a nearby event was detected — its absence means "no nearby event", so nearby is always true when the object exists.

FootTrafficExt

Real-time venue foot traffic metrics derived from FEIN audience sensing. Higher traffic = more impressions = higher value for advertisers.

{
  "current_audience": 25,
  "estimated_occupancy": 50,
  "avg_dwell_seconds": 6,
  "traffic_tier": "very_high",
  "estimated_hourly_impressions": 15000
}

IABSegmentReference

Reference table of IAB Audience Taxonomy 1.1 segment IDs used by Trillboards.

Income Segments

Lifestyle Segments

Purchase Intent Segments

Type: object

Impression

IntentExt

Purchase intent signals derived from ambient audio classification. YAMNet TFLite classifies audio SCENES (not speech content) on-device. Combined with visual shopping signals (bag detection) and demographic context to compute a composite intent score.

Higher intent_score = audience is more likely to make a purchase. DSPs can bid higher when purchase_stage is READY_TO_BUY or COMPARING.

OpenrtbRevenueScreenHourly

Revenue per screen per hour. Aggregates gross revenue from completed impressions. Refreshed from CH via clickhouseRollupSource. PG only.

PartnerReconciliationReports

Partner-reported impression and revenue counts for reconciliation against deduped completed_impressions. Ingested from VidVerto SSP API and Adipolo API / JustBaat CSV reports. PG only.

PartnerVariants

Canonical INT-PK identity for partner variants (e.g. adipolo_lower_floor_tablet). Wave 2 of pristine-state refactor. variant_name TEXT remains as the wire-format alternate-key indefinitely.

PartnerZoneCpmDaily

Per-zone per-day CPM rates from partner APIs. Used for revenue estimation and settled partner-truth allocation when partner_reconciliation_reports does not provide a newer manual/API row. PG only.

Regs

Segment

SellerDefinedSegmentReference

Seller-defined context segments in the trillboards-seller-defined data object. These are included in user.data[] with ext.segtax omitted (seller-defined taxonomy).

Venue Segments (trb_vn_*)

Geo Segments (trb_geo_*)

Traffic Tier Segments (trb_trf_*)

Dwell Time Segments (trb_dwl_*)

Daypart Segments (trb_dpt_*)

Quality Flags (trb_qly_*)

Type: object

Source

TrbNamespace

Trillboards extension sub-namespaces. Every sub-namespace is OPTIONAL and is omitted entirely when the underlying signal is unavailable for that screen — treat an absent key as "no signal", never as a zero.

This object is pinned to the request builder by tests/unit/openrtbExtNamespaceSpecParity.test.js, which derives the emitted key set from the source and fails on drift in either direction.

TrillboardsAudienceExt

Real-time audience intelligence from FEIN edge AI sensors. Updated every 10-30 seconds. These signals are MEASURED, not estimated. Unlike CTV viewability which uses probabilistic models, Trillboards uses on-device computer vision (ML Kit) and audio classification (YAMNet TFLite) to provide deterministic audience data.

{
  "attention_score": 0.87,
  "audience_count": 8,
  "crowd_density": 12,
  "ad_receptivity": 0.75,
  "emotional_engagement": 0.82,
  "mood": "happy",
  "estimated_occupancy": 25,
  "avg_dwell_ms": 4500,
  "group_composition": {
    "solo": 3
  },
  "screen_engagement": "active",
  "audience_reaction": "positive",
  "signals_age_ms": 12480
}

User

User object carrying the screen's audience segments. The trillboards-audience segments are read from the per-person records the screen keeps for the people its camera detects. Trillboards recognises the same person across screens and days and links people to device keys and mobile advertising IDs.

UserData

IAB Audience Taxonomy 1.1 segments and seller-defined context segments. Three data sources may be present in user.data[]:

1. trillboards-audience: Real-time audience segments from FEIN edge AI
2. trillboards-census: US Census ACS neighborhood demographics
3. trillboards-seller-defined: Seller-defined context segments (venue, traffic, dwell, daypart)

VASExt

Verified Attention Seconds — historical attention quality metrics for this screen. Based on 7-day rolling data from the attention_ledger.

VAS = face_count attention_score dwell_seconds. It captures the TOTAL attention delivered, not just whether someone was present. A screen with 3 attentive viewers for 10 seconds scores higher than 10 distracted viewers for 1 second.

VenueExt

Venue quality and neighborhood context for the screen's location.

Every field is omitted unless it carries positive value — a zero or unknown is dropped rather than broadcast, because a zero on the bid stream reads as "worthless" and suppresses bids. An absent field means "not known", not "scored zero".

WeatherExt

Contextual weather data for the screen's location. Cached in Redis with 60-minute TTL from WeatherAPI.com. DSPs use weather for contextual targeting (e.g., ice cream ads when hot, umbrella ads when raining).

{
  "temp_f": 85,
  "feels_like_f": 88,
  "humidity": 75,
  "condition": "Partly cloudy",
  "uv_index": 8,
  "wind_mph": 12,
  "is_raining": false,
  "temp_segment": "hot"
}